bmottin/claude_mac
1
0
Fork 0
forked from bchanot/claude
Code Pull Requests Activity

chore(memory): BDR-082 + LRN-140 + journal + CHANGELOG + TODO C1 done (seo/geo de-prescription)

Browse Source
This commit is contained in:
Bastien Chanot
2026-08-02 17:28:07 +02:00
parent c7646a9c8a
commit 325962e080
5 changed files with 42 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.claude/memory/learnings.md
+7
View File
@@ -1361,3 +1361,10 @@ rules:
- **Opus 5 traps found**: (a) Claude Code injects Opus-5-only anti-delegation prompt sections (heron_brook + subagent_steer_delegation, issue #80988; server-gated, no opt-out, absent from transcripts) — own prose stacks on top blindly; (b) NO model-default effort hold on Opus 5 — persisted effortLevel (xhigh, settings.json) silently carries over, against "start high, sweep low/medium"; run /effort sweep per model; (c) effort does NOT shorten visible output/deliverables — only prose length rules do (+30-40% docs).
- **future application**: at every model-generation bump, grep config for trait-compensating language ("counters model tendency…", "default to X") and re-verify the premise; prefer WHEN-guidance (conditions where X pays) over directional nudges — survives inversions unchanged.
- **link**: [[LRN-030]] [[BDR-081]].
## LRN-140 — de-prescription findings: dedup evaporates, self-verify is default, recall survives (2026-08-02)
- **pattern 1 — inventory dedup counts lie**: line-level inspection killed most "duplicate" pairs (seo 9 families→2 real merges; geo 7→0). Twins differ by AUDIENCE (bundle-item payload read by fresh applier vs spec rule) or MODE-RANGE (collect/judge/template/RULES) or are distinct obligations sharing a keyword (30/70 ×3 = three different rules). Dedup rule that survives: verbatim + same-audience + same-range ONLY.
- **pattern 2 — Opus 5 self-verifies unprompted**: "run it twice" instruction REMOVED → after-judge still ran score engine twice, identical output. Removing verify-prose does not remove the behavior; its value = no compounding, no contradiction burn. Confirms BDR-081 E3 mechanism, refines the payoff claim.
- **pattern 3 — de-prescription does NOT depress recall**: reworded collect caught  -encoded phone AT COLLECT (baseline collect missed it); reworded judge found new RGPD finding + self-caught false positive + corrected collect coverage claim 21/21→20/21. Integrity/honesty invariants (kept class B) carry the discipline, not the caps.
- **pattern 4 — lock strings, never shapes**: LLM-convention output layers (banners, fences, table columns, section order) wobble run-to-run in BOTH directions — baseline itself deviated from spec where after conformed (§0 ENTRIES, BUNDLE-before-SCORING). Stable contract = census-locked literal strings; anything unlocked drifts and MUST be tolerated by consumers (tier recognition "by intent" is the right pattern).
- **link**: [[BDR-082]] [[BDR-081]] [[LRN-139]] [[LRN-113]].